Zum Inhalt springen
Cashflow Hilfe
English
Esc
↑↓navigieren↵öffnen⌘Jvorschau
Auf dieser Seite

AI assistant

Connecting the optional assistant to a model of your own: what it can do, what reaches your provider, and where your token is kept.

It will only discuss this app: personal finance, how to use it, and privacy.

How do I enable it?

Open Settings → AI Assistant, switch it on, paste the API base URL (e.g. https://api.openai.com/v1), the model name (e.g. gpt-4o-mini), and your API token. Hit Test connection, then Save.

Which providers work?

Anything that speaks the OpenAI Chat-Completions API: OpenAI, OpenRouter, Groq, Together, local Ollama (http://host:11434/v1), LM Studio, vLLM, llama.cpp server, etc.

How is a connection’s availability checked?

Each connection has an Availability check setting. Fast only calls /models — quick, but a proxy such as LiteLLM answers even when the model behind it is offline. Model listed additionally requires your model id to appear in that list. Real request (default) asks the provider’s /health for the model and otherwise sends a 1-token request, so a downed backend shows up. If the proxy answers but the model does not, the badge turns amber Degraded. Settings checks all connections when you open it and then quietly every 5 minutes while the tab is visible; each badge shows when it was last checked.

What can it do?

Prefill the Add-Transaction form from a sentence like “I spent 50 at Coop on groceries, paid by credit card”; answer questions about your data (“where did I spend most last month?”) by calling read-only tools; prepare recurring rules (standing orders, subscriptions, bills) from an invoice or a description; read bank and credit-card statements you upload on the Statements screen or in the chat (PDF or a photo/screenshot) and turn them into comparable rows; and explain app features and the privacy notice. It will refuse any other topic.

Attaching files in the chat

With the paper clip, the camera or the Nextcloud button, by drag & drop or by pasting, you attach up to three files (PDF, CSV, text or image, up to 15 MB each) to your next message. An attachment is context — as if you put a document in front of the assistant. Nothing happens to it on its own.

  • If you say what to do with it (“set this up as a rule”, “import this into the UBS account”, “what’s the total?”), it does that.
  • If you send the file without text, it says briefly what the document is and proposes one thing as a question, e.g. “This is BSZ Wohnheim’s invoice for Q3, CHF 1’350. Shall I set up a quarterly rule for it?”. Your reply in the chat (“yes”, “yes, but monthly”) is the confirmation.

A PDF’s text is extracted and sent along; images only go to models that can see images. The files are kept privately in your storage: for saved conversations until you delete the conversation, in the sidebar chat for at most 7 days.

A recurring rule from an invoice

Under Settings → Recurring transactions, ✨ With AI opens a dialog: upload an invoice, premium notice or contract (or pick it from Nextcloud), optionally add a hint (“quarterly, from the UBS account”), Analyse. The AI works out the amount, due date, interval (monthly, every 2 months, quarterly, yearly …) and the billing period — also when the invoice is for the previous month or quarter — and writes a description with placeholders such as Quartal ${periodTo:Q}. It is given all your existing recurring rules as examples, so name, account and category follow your style.

The result opens in the normal rule editor, with a note on what the AI was unsure about. If the invoice belongs to an existing rule, the note offers to edit or duplicate that rule instead. Nothing is saved until you save. Without an AI connection the button is greyed out. In the chat the same works with an attached invoice or a sentence like “Netflix 15.90 monthly on the 5th from my main account”.

Which models are suitable?

For actions (prepare a transaction or rule, import a statement, attachments) the chat model must be able to call tools, and for photos it must be able to see images. The OpenAI API has no way to ask for this, so the app checks it itself with two tiny requests: after a connection is saved and via Check capabilities. The result is shown as badges on the connection. The chat prefers connections with tools; if the chosen model has none, it answers without actions and says so under the reply, and attachments are greyed out.

Can I talk to it instead of typing?

Yes, if one of your AI connections has a transcription model set (Settings → AI Assistant, e.g. whisper-1 or a local faster-whisper model). A microphone button then appears in the chat: record, stop, and the text lands in the input box for you to review before sending. Audio is recorded as a 16 kHz mono WAV, sent to that provider only, and never stored — the activity log keeps only the transcript preview.

Does it write to my data?

Hardly. The chat only prepares drafts for the Add screen and the rule editor, and the classification passes (statement lines, pending transactions) only fill suggestion fields that you accept or ignore — you always review and save manually. The one exception: it imports an attached statement (creating an import under Statements), but only when you asked for it or confirmed its proposal. All other tools are read-only.

Six tasks, one of them unprompted

The AI is used in six places. You can assign each one its own connection (and its own model) under Settings → AI Assistant:

Task When it runs
Chat when you type in the assistant
Read statement when you upload a statement
Categorise lines right after reading, for lines without a match
Suggestions for pending transactions automatically, once for every line that arrives
Voice input when you press the microphone
Recurring rule from invoice when you press ✨ With AI on the recurring rules

If one provider serves several models (a LiteLLM proxy or an Ollama host, say), you do not need a second connection: add them to the connection under Additional models — type the name or pick it from the provider’s model list — and save. Once a task is bound to that connection, Model for this action lets you choose the default model or one of the additional ones, followed by the list the provider itself reports. Availability is still checked once per connection, using the default model.

The fourth is the only one that starts without you doing anything — for instance when FinReader reports a payment. If you do not want that, assign it no connection or leave it switched off.

What is sent to my provider?

More than just your messages. Almost every request carries an overview of your ledger: accounts with balances, categories with budget and month-to-date figures, your tags, frequent descriptions — and the most recent transactions in full, with date, amount, description, account and category. Without this context the model can neither suggest a category nor answer a question about your money.

How much of it goes along is set per connection:

Level Period Transactions
Off — none
Compact 30 days up to 400
Full 30 days up to 400, more detail
XL 180 days up to 1,200

Two exceptions you should know about:

  • Suggestions for pending transactions upgrade Off to Compact. Without a category list the pass cannot suggest anything — so Off does not prevent anything being sent here. The line’s original notification text is sent along too.
  • Categorise lines always sends the Full level and converts to CHF while doing so, regardless of your setting.

The complete breakdown per task is in the privacy notice, section 6a.

What is in the AI log?

Under Settings → AI log every request is recorded: time, connection, model, provider host, duration, token usage — and excerpts of the content: the beginning of your last message, the beginning of the reply, the arguments and results of each tool call, and for voice input the beginning of the transcript. Your API token is never written down there.

This log is not deleted automatically. Unlike the general audit log, which is cleared out after a year, it stays until you empty it yourself — the Clear contents button sits above the list. It removes the excerpts; time, kind, connection, model, duration and token counts stay for AI usage.

Above the list you filter by kind, connection and model.

How many tokens do I use?

Settings → AI usage → Show AI usage opens a page of its own with the usage from the AI log.

  • At the top you pick the period: This month, Last month, This year or your own Date range.
  • The figures: requests (and in how many chats), tokens in (of them cached, where the provider reports it), tokens out, total, failed requests, and requests not reported — where the provider sent no token count. They are counted, never estimated.
  • The chart shows tokens per day, or per month for a year. It groups by Connection, Model or Feature (the six tasks above). Clicking an entry in the legend hides that group, and the axis adjusts.
  • The table (one card per line on a phone) shows requests, tokens in and out, total, share, Ø tokens per request, Ø duration and errors. A connection opens into its models. Log → opens the AI log filtered to that line.
  • “n as fallback” means that many requests ran on this connection although the task is assigned to another one, because that one was not reachable. “no longer set up” marks a model the connection no longer offers.

The assistant’s tool calls are not model requests and are not counted; failed requests are, because providers usually bill them anyway. Entries from before the connection was recorded are matched by host and model; for them a fallback cannot be told.

Where is my API token stored?

Server-side in the ai_endpoints table, one per connection. It is not returned to the browser, but the server operator can read it — treat it like other credentials on this instance.

War diese Seite hilfreich?